对张量流图的一部分进行基准测试的正确方法是什么？

Question

我想对图形的某些部分进行基准测试，为了简单起见，我使用conv_block的只是 conv3x3。

1. x_np循环中使用的是否相同或者我每次都需要重新生成它？
2. 在运行实际基准测试之前，我是否需要进行一些“热身”运行（似乎这是 GPU 基准测试所需要的）？如何正确地做到这一点？是sess.run(tf.global_variables_initializer())够了吗？
3. 在python中测量时间的正确方法是什么，即更精确的方法。
4. 在运行脚本之前，我是否需要在 linux 上重置一些系统缓存（也许禁用 np.random.seed 就足够了）？

示例代码：

import os
import time

import numpy as np
import tensorflow as tf

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)

np.random.seed(2020)


def conv_block(x, kernel_size=3):
    # Define some part of graph here

    bs, h, w, c = x.shape
    in_channels = c
    out_channels = c

    with tf.variable_scope('var_scope'):
        w_0 = tf.get_variable('w_0', [kernel_size, kernel_size, in_channels, out_channels], initializer=tf.contrib.layers.xavier_initializer())
        x = tf.nn.conv2d(x, w_0, [1, 1, 1, 1], 'SAME')

    return x


def get_data_batch(spatial_size, n_channels):
    bs = 1
    h = spatial_size
    w = spatial_size
    c = n_channels

    x_np = np.random.rand(bs, h, w, c)
    x_np = x_np.astype(np.float32)
    #print('x_np.shape', x_np.shape)

    return x_np


def run_graph_part(f_name, spatial_size, n_channels, n_iter=100):
    print('=' * 60)
    print(f_name.__name__)

    tf.reset_default_graph()
    with tf.Session() as sess:
        x_tf = tf.placeholder(tf.float32, [1, spatial_size, spatial_size, n_channels], name='input')
        z_tf = f_name(x_tf)
        sess.run(tf.global_variables_initializer())

        x_np = get_data_batch(spatial_size, n_channels)
        start_time = time.time()
        for _ in range(n_iter):
            z_np = sess.run(fetches=[z_tf], feed_dict={x_tf: x_np})[0]
        avr_time = (time.time() - start_time) / n_iter
        print('z_np.shape', z_np.shape)
        print('avr_time', round(avr_time, 3))

        n_total_params = 0
        for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='var_scope'):
            n_total_params += np.prod(v.get_shape().as_list())
        print('Number of parameters:', format(n_total_params, ',d'))


if __name__ == '__main__':
    run_graph_part(conv_block, spatial_size=128, n_channels=32, n_iter=100)

Answer 1

回答您的主要问题“对张量流图的部分进行基准测试的正确方法是什么？”：

Tensorflow 包含一个抽象类，为张量流基准测试提供帮助器：Benchmark。

因此，Benchmark可以创建一个对象并用于在张量流图的一部分上执行基准测试。在下面的代码中，实例化了一个基准对象，然后run_op_benchmark调用该方法。run_op_benchmark传递给会话、conv_block张量（在本例中）、 a feed_dict、烧录迭代次数、所需的最小迭代次数、一个布尔标志（以防止基准测试也计算内存使用情况）和一个方便的名称。该方法返回一个包含基准测试结果的字典：

benchmark = tf.test.Benchmark()
results = benchmark.run_op_benchmark(sess=sess, op_or_tensor=z_tf, 
                                     feed_dict={x_tf: x_np}, burn_iters=2, 
                                     min_iters=n_iter, 
                                     store_memory_usage=False, name='example')

可以将此代码块插入到您的代码中，如下所示以比较两个基准测试：

import os
import time

import numpy as np
import tensorflow as tf

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)

np.random.seed(2020)


def conv_block(x, kernel_size=3):
    # Define some part of graph here

    bs, h, w, c = x.shape
    in_channels = c
    out_channels = c

    with tf.compat.v1.variable_scope('var_scope'):
        w_0 = tf.get_variable('w_0', [kernel_size, kernel_size, in_channels, out_channels], initializer=tf.keras.initializers.glorot_normal())
        x = tf.nn.conv2d(x, w_0, [1, 1, 1, 1], 'SAME')

    return x


def get_data_batch(spatial_size, n_channels):
    bs = 1
    h = spatial_size
    w = spatial_size
    c = n_channels

    x_np = np.random.rand(bs, h, w, c)
    x_np = x_np.astype(np.float32)
    #print('x_np.shape', x_np.shape)

    return x_np


def run_graph_part(f_name, spatial_size, n_channels, n_iter=100):
    print('=' * 60)
    print(f_name.__name__)

    tf.reset_default_graph()
    with tf.Session() as sess:
        x_tf = tf.placeholder(tf.float32, [1, spatial_size, spatial_size, n_channels], name='input')
        z_tf = f_name(x_tf)
        sess.run(tf.global_variables_initializer())

        x_np = get_data_batch(spatial_size, n_channels)
        start_time = time.time()
        for _ in range(n_iter):
            z_np = sess.run(fetches=[z_tf], feed_dict={x_tf: x_np})[0]
        avr_time = (time.time() - start_time) / n_iter
        print('z_np.shape', z_np.shape)
        print('avr_time', round(avr_time, 3))

        n_total_params = 0
        for v in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='var_scope'):
            n_total_params += np.prod(v.get_shape().as_list())
        print('Number of parameters:', format(n_total_params, ',d'))

        # USING TENSORFLOW BENCHMARK
        benchmark = tf.test.Benchmark()
        results = benchmark.run_op_benchmark(sess=sess, op_or_tensor=z_tf, 
                                             feed_dict={x_tf: x_np}, burn_iters=2, min_iters=n_iter,
                                             store_memory_usage=False, name='example')

        return results


if __name__ == '__main__':
    results = run_graph_part(conv_block, spatial_size=128, n_channels=32, n_iter=100)

张量流库本身中的基准测试类的实现提供了有关其他问题答案的提示。由于张量流实现不需要feed_dict为每个基准迭代使用新的，因此问题 1）的答案似乎是“x_np循环中使用的是否可以相同，或者我每次都需要重新生成它？” 是可以使用相同的x_np每个循环。关于问题2），看来确实有必要进行一些“热身”。Tensorflow 库实现建议的默认烧录迭代次数为 2。关于问题 3），timeit它是测量小代码片段执行时间的出色工具。但是，tensorflow 库本身的使用time.time()方式与您所做的类似：run_op_benchmark(source)。有趣的是，张量流基准测试实现报告的是操作间隔时间的中值而不是平均值（大概是为了使基准测试对异常值更加稳健）。

Answer 2

添加到 Steve's answer的令人敬畏的解释中，以下内容在 TensorFlow-GPU v2.3 上对我有用

import tensorflow as tf

tf.config.experimental.set_memory_growth(tf.config.experimental.list_physical_devices('GPU')[0], True)

import os
import time

import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)

np.random.seed(2020)



def conv_block(x, kernel_size=3):
    # Define some part of graph here

    bs, h, w, c = x.shape
    in_channels = c
    out_channels = c

    with tf.compat.v1.variable_scope('var_scope'):
        w_0 = tf.compat.v1.get_variable('w_0', [kernel_size, kernel_size, in_channels, out_channels], initializer=tf.keras.initializers.glorot_normal())
        x = tf.nn.conv2d(x, w_0, [1, 1, 1, 1], 'SAME')

    return x


def get_data_batch(spatial_size, n_channels):
    bs = 1
    h = spatial_size
    w = spatial_size
    c = n_channels

    x_np = np.random.rand(bs, h, w, c)
    x_np = x_np.astype(np.float32)
    #print('x_np.shape', x_np.shape)

    return x_np


def run_graph_part(f_name, spatial_size, n_channels, n_iter=100):
    print('=' * 60)
    print(f_name.__name__)

#     tf.reset_default_graph()
    tf.compat.v1.reset_default_graph()
    
    
    with tf.compat.v1.Session() as sess:
        x_tf = tf.compat.v1.placeholder(tf.float32, [1, spatial_size, spatial_size, n_channels], name='input')
        z_tf = f_name(x_tf)
        
        sess.run(tf.compat.v1.global_variables_initializer())

        x_np = get_data_batch(spatial_size, n_channels)
        
        start_time = time.time()
        
        for _ in range(n_iter):
            z_np = sess.run(fetches=[z_tf], feed_dict={x_tf: x_np})[0]
        avr_time = (time.time() - start_time) / n_iter
        
        print('z_np.shape', z_np.shape)
        print('avr_time', round(avr_time, 3))

        n_total_params = 0
        
        for v in tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope='var_scope'):
            n_total_params += np.prod(v.get_shape().as_list())
        
        print('Number of parameters:', format(n_total_params, ',d'))

        # USING TENSORFLOW BENCHMARK
        benchmark = tf.test.Benchmark()
        results = benchmark.run_op_benchmark(sess=sess, op_or_tensor=z_tf, 
                                             feed_dict={x_tf: x_np}, burn_iters=2, min_iters=n_iter,
                                             store_memory_usage=False, name='example')

        return results


if __name__ == '__main__':
    results = run_graph_part(conv_block, spatial_size=512, n_channels=32, n_iter=100)

在我的情况下会输出类似 -

============================================================
conv_block
z_np.shape (1, 512, 512, 32)
avr_time 0.072
Number of parameters: 9,216
entry {
  name: "TensorFlowBenchmark.example"
  iters: 100
  wall_time: 0.049364686012268066
}